Multi-agent cognitive self-evolution method and system

Abstract

The application relates to the technical field of computer software, and discloses a multi-agent cognitive self-evolution method and system; the method comprises the following steps: broadcasting a macro event stream as a global background to all agents, and allowing the agents to select the most relevant news topics from the news topic stream; the agents perform cognitive processing based on received double-stream data, a character portrait and dynamic memory, and output attitude scores and behavior responses; the agents are guided to perform dynamic interaction based on content similarity and timeliness in an information-driven group, and the dynamic memory is updated; attitude and behavior response data of all agents are aggregated to synthesize a macro trend, the updated dynamic memory after the current round is taken as the dynamic memory of the agents in the next round, and double-stream data in the next round is loaded; the application provides a high-fidelity and dynamically adaptive computing experiment platform for simulating group attitude evolution in a complex dynamic environment and reproducing social historical trends.

Description

Technical Field

[0001] This invention relates to the field of computer software technology, and specifically to a multi-agent cognitive self-evolution method and system. Background Technology

[0002] Multi-agent cognitive self-evolutionary systems have significant application value in the following aspects: 1) Risk identification and early warning: The simulation system can monitor changes in group emotions and the degree of event evolution, providing intelligent early warning solutions. 2) Revealing the mechanism of event propagation and socio-psychological evolution: By simulating the interactive behavior of a large number of heterogeneous agents under multi-source social events, the system analyzes the formation mechanism of complex social phenomena and reveals the nonlinear evolution law of micro-individual cognitive interaction emerging into macro-social forms. 3) Cognitive intervention and evolutionary governance: Constructing a tracking and intervention verification mechanism for cognitive change trajectories provides an experimental platform for event evolution analysis, consensus building strategies, etc., supporting the precision and controllability of social intelligent decision-making.

[0003] Previous agent-based models (ABMs) were designed to simulate macro-level social dynamics through collective interactions between individuals. However, these models typically rely on simple heuristic rules and fail to fully reflect individual diversity and cognitive reasoning abilities. In contrast, agents based on large language models (LLMs) exhibit advanced semantic understanding and decision-making capabilities, able to interpret complex environmental information and thus simulate individuals' deep cognition of the environment. The design of data injection methods and interaction rules has become a core issue in research on large language model-driven social simulators. For example, Liu Yuhan et al. focused on the propagation mechanism of events, utilizing the human-like capabilities of large models in perception and reasoning to construct a static social network structure. They reproduced the evolution path of events within the group and the accumulation process of cognitive biases, achieving a preliminary exploration of social interaction; Nicholas Sukiennik et al. constructed a small-scale information environment aligned with real time, and by aggregating individual intentions, they inferred the evolutionary trend of social groups towards specific long-term events, to some extent reproducing the dynamic trajectory of group attitudes over time.

[0004] Despite significant progress in social simulation, existing methods still face limitations in constructing large-scale, long-term social evolution simulation systems highly aligned with real-world history, particularly in the dynamic coupling of individual cognition with the multi-source information environment. Most current methods rely on pre-defined static interaction topologies or single-dimensional static data injection, neglecting the dynamic restructuring of social relations during social evolution and the continuous driving force of multi-source real-time information flow on group cognition. On one hand, at the data-driven level, existing studies typically employ one-time initialization or a single information source. However, the evolution of social systems is essentially an open system constantly exchanging information with its external environment; group cognition is often influenced by both macro-level events and micro-level news topics. Single or static data input leads to information homogenization or contextual gaps in the simulation environment, making it difficult to reproduce the true evolutionary trajectory of group attitudes under complex environments, arising from the interweaving of multi-source information. On the other hand, in the dynamic interaction process, existing methods largely rely on pre-defined fixed interaction rules, severing the causal link between information selection and social interaction. In real-world social evolution, individuals' interaction partners are not static but exhibit significant news-topic-driven characteristics. That is, event dissemination circles temporarily form based on current hot news topics or shared concerns and then dynamically reorganize. Pre-defined static topologies cannot simulate the fluidity of social relationships within the event dissemination field. Therefore, constructing a social simulation mechanism that can integrate real-time, multi-source, heterogeneous event streams and support agents in autonomously adjusting their interaction partners based on cognitive conflict and content resonance has become a crucial path to accurately aligning with the real-world event dissemination landscape. In conclusion, modeling multi-source data-driven dynamic agent interactions is a significant challenge that current multi-agent simulation systems urgently need to address. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multi-agent cognitive self-evolution method and system.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a multi-agent cognitive self-evolution method, comprising the following steps: S1 initializes the global environment and agent group based on real-world data, and configures a person profile and dynamic memory for each agent; S2 constructs environmental information through dual-stream data consisting of macro event streams and news topic streams. It broadcasts the macro event streams as a global background to all agents and allows each agent to select the most relevant news topics from the news topic stream based on its own profile and dynamic memory. S3: The agent performs cognitive processing based on the received dual-stream data, person profile, and dynamic memory, and outputs attitude scores and behavioral responses. S4, based on behavioral response, dynamically clusters agents that are interested in the same news topic into temporary information-driven groups. Within the information-driven groups, agents are guided to interact dynamically based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memory and update dynamic memory. S5 aggregates the attitude scores and behavioral response data of all agents to synthesize the macro situation, uses the dynamic memory updated in the current round as the agent dynamic memory for the next round, and loads the dual-stream data for the next round. S6. Repeat steps S2 to S5 until the preset simulation termination condition is met.

[0007] In one embodiment, allowing each agent to select the most relevant news topics from the news topic stream based on its own persona and dynamic memory specifically includes: Constructing a candidate information pool based on news topic streams This enables the agent to react to macroscopic events in round t. Self-portrait And the dynamic memory of round t-1 It independently calculates and selects the most relevant news topics. : ; in, To select the engine function, the semantic alignment between the input news topics and the agent's persona, as well as dynamic memory, is comprehensively evaluated to retrieve the K news topics with the highest semantic alignment. This represents a large model.

[0008] In one embodiment, the agent performs cognitive processing based on the received dual-stream data, user profile, and dynamic memory, and outputs an attitude score and behavioral response, specifically including: Introducing a decision engine based on prompt words. With self-reflective memory operators In the t-th round of interaction, the i-th agent Attitude scores are generated based on input information. Behavioral response and the updated dynamic memory It will be made by a large model The decision-making process of a driven intelligent agent is modeled as follows: ; For the i-th agent, select the news topic that interests them most in round t. Create a portrait of the i-th intelligent agent. For the i-th agent, the dynamic memory in round t-1; The reference information flow is constructed by the i-th agent in the t-th round based on the visual reply content and dynamic information flow of the k-th news topic; The input information is composed of the above; The attitude score is used to quantify the agent's emotional polarity and emotional intensity towards current macro events and specific news topics; The behavioral response refers to the social response actions taken by the agent, including posting opinions, forwarding, or replying.

[0009] In one embodiment, guiding agents to dynamically interact within an information-driven group based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memories and update dynamic memories, specifically includes: Based on visual response content With dynamic information flow Constructing a reference information flow ;

[0010] Information-driven group for the k-th news topic in the t-th round of interaction In the process, calculate the i-th agent in the t-th round of interaction. With the i-th intelligent agent Recommended scores between For intelligent agents Sort the other agents in descending order of their recommendation scores and select the one with the highest recommendation score. individual agents and Enable dynamic interaction: ; in, Represents cosine similarity. The attenuation rate; Indicates the timestamp of the action's publication. This represents the action taken by the i-th agent in the t-th round of interaction regarding the current news topic. This represents the behavioral response of the j-th agent to the current news topic in the t-th round of interaction; Dynamic memory is updated based on reflective memory: ; This represents the dynamic memory of the i-th agent in round t. The prompt words are used to guide the large model in updating its memory. express After After processing and Dynamic memory updated jointly.

[0011] In one embodiment, the reference information stream The construction process specifically includes: Visual reply content This represents what the i-th agent can see in round t, targeting The set of observable responses: ; in, This represents the k-th news topic of interest selected by the i-th agent in the t-th round. This indicates that for the same news topic, the j-th agent... Execution with the i-th intelligent agent The same responsive action; Represents the i-th intelligent agent In the t round, targeting the A news item is a dynamic information stream received from other agents, which is obtained by aggregating all information streams from agents within the same group that perform the same operation driven by the same information: .

[0012] Secondly, the present invention provides a multi-agent cognitive self-evolutionary system, comprising the following steps: The initialization module initializes the global environment and agent group based on real-world data, and configures a person profile and dynamic memory for each agent. The dual-stream data integration module constructs environmental information through dual-stream data consisting of macro event streams and news topic streams. It broadcasts the macro event streams as a global background to all intelligent agents and allows each intelligent agent to select the most relevant news topics from the news topic stream based on its own profile and dynamic memory. In the initial behavior decision-making module, the agent performs cognitive processing based on the received dual-stream data, user profile, and dynamic memory, and outputs attitude scores and behavioral responses. The interactive reflection module, based on behavioral response, dynamically clusters agents that are interested in the same news topic into temporary information-driven groups. Within the information-driven groups, agents are guided to interact dynamically based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memories and update dynamic memories. The situation synthesis module aggregates the attitude scores and behavioral response data of all agents to synthesize the macro situation, uses the dynamic memory updated in the current round as the agent dynamic memory for the next round, and loads the dual-stream data for the next round.

[0013] In one embodiment, allowing each agent to select the most relevant news topics from the news topic stream based on its own persona and dynamic memory specifically includes: Constructing a candidate information pool based on news topic streams This enables the agent to react to macroscopic events in round t. Self-portrait And the dynamic memory of round t-1 It independently calculates and selects the most relevant news topics. : ; in, To select the engine function, the semantic alignment between the input news topics and the agent's persona, as well as dynamic memory, is comprehensively evaluated to retrieve the K news topics with the highest semantic alignment. This represents a large model.

[0014] In one embodiment, the agent performs cognitive processing based on the received dual-stream data, user profile, and dynamic memory, and outputs an attitude score and behavioral response, specifically including: Introducing a decision engine based on prompt words. With self-reflective memory operators In the t-th round of interaction, the i-th agent Attitude scores are generated based on input information. Behavioral response and the updated dynamic memory It will be made by a large model The decision-making process of a driven intelligent agent is modeled as follows: ; For the i-th agent, select the news topic that interests them most in round t. Create a portrait of the i-th intelligent agent. For the i-th agent, the dynamic memory in round t-1; The reference information flow is constructed by the i-th agent in the t-th round based on the visual reply content and dynamic information flow of the k-th news topic; The input information is composed of the above; The attitude score is used to quantify the agent's emotional polarity and emotional intensity towards current macro events and specific news topics; The behavioral response refers to the social response actions taken by the agent, including posting opinions, forwarding, or replying.

[0015] In one embodiment, guiding agents to dynamically interact within an information-driven group based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memories and update dynamic memories, specifically includes: Based on visual response content With dynamic information flow Constructing a reference information flow ;

[0016] Information-driven group for the k-th news topic in the t-th round of interaction In the process, calculate the i-th agent in the t-th round of interaction. With the i-th intelligent agent Recommended scores between For intelligent agents Sort the other agents in descending order of their recommendation scores and select the one with the highest recommendation score. individual agents and Enable dynamic interaction: ; in, Represents cosine similarity. The attenuation rate; Indicates the timestamp of the action's publication. This represents the action taken by the i-th agent in the t-th round of interaction regarding the current news topic. This represents the behavioral response of the j-th agent to the current news topic in the t-th round of interaction; Dynamic memory is updated based on reflective memory: ; This represents the dynamic memory of the i-th agent in round t. The prompt words are used to guide the large model in updating its memory. express After After processing and Dynamic memory updated jointly.

[0017] In one embodiment, the reference information stream The construction process specifically includes: Visual reply content This represents what the i-th agent can see in round t, targeting The set of observable responses: ; in, This represents the k-th news topic of interest selected by the i-th agent in the t-th round. This indicates that for the same news topic, the j-th agent... Execution with the i-th intelligent agent The same responsive action; Represents the i-th intelligent agent In the t round, targeting the A news item is a dynamic information stream received from other agents, which is obtained by aggregating all information streams from agents within the same group that perform the same operation driven by the same information: .

[0018] The system and method in this invention correspond to each other; the specific technical solutions applicable to the method are also applicable to the system.

[0019] Compared with the prior art, the beneficial technical effects of the present invention are: The historical alignment of social situation simulation is significantly improved: through a dual-stream data integration engine, this invention achieves precise coupling between macro-historical processes and micro-individual cognition. This mechanism allows agents to be guided by macro-events globally while also possessing the ability to personalize the selection of massive news topics, effectively overcoming the information noise or detachment from reality caused by the limited data in traditional methods. This strategy significantly reduces the deviation from real history in long-term simulations and exhibits strong robustness as the agent scales up, enabling it to realistically and stably reproduce the evolutionary trajectory of group attitudes in complex environments.

[0020] Superior to static networks in dynamic event emergence and adaptive interaction capabilities: Compared with static networks, dynamic social groups can effectively capture the dynamic evolution of group attitudes. Through a continuous learning closed-loop mechanism of perception-response-adjustment, agents can update their attitudes in real time according to changes in intra-group interactions, improving their responsiveness to hotspot evolution, emotional fluctuations and group polarization. This effectively supports the evolutionary modeling of complex individual behaviors in multi-agent social simulations and provides more realistic explanatory technical support for event inference and intervention assessment.

[0021] In summary, this invention provides a computational experimental platform with high fidelity and dynamic adaptability for simulating the evolution of group attitudes in complex dynamic environments and reproducing macro-social historical trends. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0023] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown, a multi-agent cognitive self-evolution method includes the following steps: S1 initializes the global environment and agent group based on real-world data, and configures a person profile and dynamic memory for each agent; S2 constructs environmental information through dual-stream data consisting of macro event streams and news topic streams. It broadcasts the macro event streams as a global background to all agents and allows each agent to select the most relevant news topics from the news topic stream based on its own profile and dynamic memory. S3: The agent performs cognitive processing based on the received dual-stream data, person profile, and dynamic memory, and outputs attitude scores and behavioral responses. S4, based on behavioral response, dynamically clusters agents that are interested in the same news topic into temporary information-driven groups. Within the information-driven groups, agents are guided to interact dynamically based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memory and update dynamic memory. S5 aggregates the attitude scores and behavioral response data of all agents to synthesize the macro situation, uses the dynamic memory updated in the current round as the agent dynamic memory for the next round, and loads the dual-stream data for the next round. S6. Repeat steps S2 to S5 until the preset simulation termination condition is met.

[0025] To overcome the limitations of existing social simulation methods in terms of reliance on static interaction rules and data timeliness, this invention proposes a multi-agent cognitive self-evolution method driven by multi-source data. This method integrates macro-historical processes with micro-individual cognition to construct a simulation environment dynamically aligned with the real world. Within this framework, the invention innovatively introduces a dual-stream data integration engine, driving the simulation by processing macro-level event streams and micro-level news topics in parallel. This allows the system to not only establish a global historical context using macro-events but also supports agents in autonomously ingesting personalized news topics based on their own profiles and interests, thus balancing the accuracy of macro trends with the diversity of micro-behaviors in large-scale social simulations. Furthermore, this invention designs a content-aware interaction module oriented towards dynamic news topics. This module constructs an interaction group for agents oriented towards dynamic news topics, capable of identifying and aggregating agents interested in the same news topic, dynamically dividing them into different interest groups for intra-group interaction. This enables agents to break through the limitations of preset static networks, autonomously reorganizing social connections based on the real-time information environment, thereby realistically simulating the dynamic process of macro-social relationship evolution driven from the bottom up by micro-individual interactions.

[0026] The present invention will be described in detail below in several parts.

[0027] 1. Global environment construction and agent swarm initialization.

[0028] In response to the task request for social situation simulation, the simulator first constructs a simulation environment based on real-world data and initializes a multi-agent group. Based on the socio-demographic distribution characteristics of the target simulated group, the simulator generates agent instances and configures two core attributes for each agent: 1) Multi-dimensional persona: covering attributes such as age, occupation, and personality traits, aiming to provide the agent with a heterogeneous cognitive foundation, ensuring that it can make reasonable personalized responses and behavioral decisions based on specific roles; 2) Dynamic memory: used to persistently store individual experiences and interaction history during the simulation process, serving as input for the agent's cognitive updates and continuous decision-making in the next stage.

[0029] 2. Dual-stream data integration engine builds environment information.

[0030] To ensure precise alignment between macro-historical trends and micro-individual cognition in the simulation, a dual-stream data integration engine constructs parallel information channels to inject dynamic environmental data into the agents, achieving accurate synchronization between the simulated environment and real-world historical processes. First, for macro-level events, the simulator broadcasts real-time macro-level events as global background information to all agents in the simulation environment. This establishes a macro-level benchmark for the simulation process, preventing collective cognition from deviating from the historical timeline. Second, based on the received macro-level background, the simulator activates a personalized autonomous ingestion module. This module empowers agents to proactively retrieve information based on cognitive preferences—that is, based on their own profiles and historical memories, they autonomously select and ingest the news topics of greatest interest from daily real-time information, achieving a unification of macro-level common constraints and micro-level personalized input.

[0031] In a preferred embodiment, the present invention designs a method for real-time updating of environmental information based on dynamic information flow. This method is used to solve the problem of dynamic matching between macro-level events and micro-level individual interests. Instead of random distribution, this mechanism constructs a candidate information pool using daily information. This enables intelligent agents to react to current macroscopic events. Self-portrait and recent memory state It independently calculates and selects the most relevant news topics. This autonomous uptake process is defined as: ; in, As a selection engine function, the K news topics with the highest semantic alignment are retrieved by comprehensively measuring the semantic alignment between the input news topics and the agent's profile and dynamic memory. This mechanism ensures that the dual-stream data at the input end is synchronized with historical reality in the time dimension and conforms to the agent's personalized preferences in the cognitive dimension, providing accurate contextual input for subsequent cognitive reasoning.

[0032] 3. Agent cognitive processing and initial behavioral decision-making.

[0033] Upon receiving dual-stream data, the agent, driven by a large language model, interprets and makes decisions based on its constructed cognition of external information. The agent treats received macro-level events and daily information as environmental stimuli, combining these with its long-term, fixed persona and short-term accumulated historical memories to perform logical reasoning and sentiment analysis. At this stage, the agent's decision output includes two core dimensions: 1) Quantitative assessment of attitude scores: The agent outputs a continuous attitude score, quantifying its emotional polarity and support strength towards the current external environment and specific news topics; 2) Behavioral response decisions: The agent autonomously decides on specific social response actions to particular information, including publishing opinions, forwarding, or replying. This decision marks the agent's transformation from a cognitive state to a social behavior state, providing a foundational signal for subsequent group interactions.

[0034] In agent behavior decision-making, an individual's cognitive state continuously evolves through interaction. This invention introduces a decision engine. With self-reflective memory operators This constructs a closed-loop mechanism of "perception-action-reflection." In each round of interaction, the agent generates an attitude score based on the input information. Behavioral response and the updated dynamic memory The decision-making process of an agent driven by a large model is modeled as follows: ; in, This is a reference information stream, containing visual reply content. With dynamic information flow .

[0035] Decision Engine Based on prompt words, it guides the large model in decision-making. Self-reflective memory operator. Based on the prompt words, the large model is used to update dynamic memory.

[0036] 4. News-driven dynamic interaction and memory self-reflection.

[0037] To simulate group interactions and opinion evolution based on news topic resonance in real society, agents can interact through an information-driven dynamic interaction mechanism. This process includes three stages: dynamic group construction, intra-group interaction execution, and memory update: 1) Dynamic grouping based on news topics: Aggregating the behavioral decisions made by all agents in the simulator, agents who have the willingness to interact on the same news topic are clustered into a temporary information-driven group. This mechanism breaks the limitations of traditional static networks and constructs a dynamic local interaction topology based on the current focus of interest. 2) Within the information-driven group, agents do not interact randomly, but select interaction objects based on reference information flow. Agents prioritize in-depth interaction (such as replying to comments) with neighbors who have similar opinions and high activity levels. 3) Self-reflective memory update: After completing a round of interaction, the agent triggers a self-reflection mechanism. The agent summarizes and reflects on the current interaction experience, including the opinions of others and its own interactive behavior, generating a self-reflective memory. This memory is written into the agent's dynamic memory to update its belief system, ensuring that the agent can make coherent decisions based on the latest social experience in the next round of simulation.

[0038] This invention constructs a reference information stream. This makes it possible for information to drive the behavior of other users within the group, which is influenced by visual response content. With recommendation-based dynamic information flow Joint composition. Its construction logic is formally expressed as: ; ; ; This indicates that for the same news topic, the j-th agent... Execution with the i-th intelligent agent The same reflexive action, which includes actions such as replying, forwarding, and posting.

[0039] By aggregating all information streams from agents performing the same operation within the same interaction group, we obtain: .

[0040] For the i-th agent in the t-th round of interaction With the i-th intelligent agent Recommended scores between: ; Cosine similarity can be represented using the sentence transformer function, which measures the similarity of content published between agents. The timeliness factor corresponds to the decay rate. .

[0041] Sort the agents in descending order of their recommendation scores and select the one with the highest recommendation score. The intelligent agents interact dynamically.

[0042] Dynamic memory is updated based on reflective memory: ; This represents the dynamic memory of the i-th agent in round t. The prompt words are used to guide the large model in updating its memory. express After After processing and Dynamic memory updated jointly.

[0043] 5. Macro-level situation synthesis and simulation iterative cycle.

[0044] After completing the micro-interactions and cognitive updates for the current simulation round, the simulator performs a global state summary and time step progression. First, the simulator aggregates the attitude scores and behavioral response data of all agents, mapping discrete individual micro-behaviors to continuous macro-social evolutionary trends. Then, upon entering the next simulation round, agents utilize updated dynamic memories, combined with new real-time macro-event flows and news topic flows as the basis for decision-making. Through this continuous iterative cycle of "event injection-cognitive processing-dynamic interaction," the simulator can generate temporally coherent social evolutionary trajectories, thereby revealing how micro-individual interactions emerge from the bottom up to generate macro-social characteristics.

[0045] This invention develops a dual-stream data integration engine and designs a collaborative mechanism for the autonomous ingestion of macro-level events and micro-level news. It utilizes a semantic alignment algorithm to accurately match news topics that match the agent's profile from a massive information pool. This engine balances the macro-level consistency of historical processes with the micro-level differences in individual cognition, solving the problems of data lag and information homogenization in traditional simulations, and ensuring a high degree of synchronization between the simulation environment and real history.

[0046] This invention designs a multi-source data-driven dynamic interaction module and a recommendation-score-based dynamic interaction method for event propagation. By quantifying the alignment of behavioral content and the timeliness decay factor, it guides the agent to autonomously select highly relevant neighbors for interaction within dynamic news topic groups. This method overcomes the limitations of static network topology and realistically simulates the "homogeneous connection" and relationship reorganization process based on news topic resonance in real-world social networks.

[0047] This invention constructs a cognitive evolution loop of "perception-action-reflection," introducing a self-reflective memory module to quantify the impact of interactive feedback on the belief system, enabling the agent to internalize instantaneous social interactions into long-term experiential memory. Through this continuous cognitive update at the micro level, a bottom-up emergence from individual interactions to macro-social trends is achieved.

[0048] Example: This embodiment uses the simulation of a specific social event as an example, setting 20 simulation rounds, with each round having a quarterly time interval. The number of agents is set to... (This invention supports scaling to a larger scale), the simulator processed more than 70,000 interactions between agents and information during the simulation.

[0049] Step 1: Upon receiving a social situation simulation task request, the simulator first loads a specific event dataset, constructing a dynamic information database strictly aligned with the real world. Simultaneously, the simulator initializes the agent group using a large model, assigning a multi-dimensional profile to each agent based on real demographic data. Specifically, the simulator employs stratified sampling or probability distribution matching to generate an attribute set containing name, age (e.g., based on a truncated normal distribution), gender, occupation, educational background, and personality traits, to more accurately reflect the demographic structure of the target society and ensure cognitive diversity within the group. Furthermore, to simulate initial disagreements and cognitive biases in group viewpoints, each agent has an initial attitude score based on a specific event, ranging from -10 to 10. Agents also possess dynamic memory, storing and managing their individual experiences and interaction history throughout the simulation cycle, laying the foundation for dynamic decision-making in subsequent simulations. This multi-layered initialization provides agents with a solid micro-cognitive foundation, supporting personalized decision-making and dynamic interactions driven by specific events, while ensuring a deep reproduction of cognitive heterogeneity at the individual level and achieving a high degree of alignment between the macro-evolutionary trajectory and real-world history.

[0050] Step 2: To ensure historical alignment and individual heterogeneity during simulation, the dual-stream data integration engine injects information into the agents through two parallel channels, macro and micro, to align the macro-historical context with the micro-individual cognition: 1) Macro-event broadcasting: Real-time macro-event sets (including occurrence time, title, and detailed content) are broadcast to all agents, establishing the global historical context of the simulation. 2) Personalized agent ingestion: The large model-driven agent combines its own profile and dynamic memory to select information of personal preference from the news topic database based on the current context. This step ensures that the information received by the agent is both consistent with historical facts on a macro level and reflects individual interest differences on a micro level, avoiding the problem of information homogenization.

[0051] Step 3: The agent makes multi-dimensional cognitive decisions based on person profiles, dynamic memory mechanisms, and reference information streams. First, person profiles provide the agent with a stable baseline of stance and a personalized background, ensuring that it exhibits cognitive tendencies consistent with its identity when facing different issues. Second, dynamic memory mechanisms enable the agent to adjust its current strategy based on past interaction history and accumulated experience, maintaining long-term cognitive coherence through reflection mechanisms and avoiding inconsistencies in decision-making. Finally, reference information streams provide the current social context, including visible neighbor responses and highly relevant dynamics recommended by the system, providing immediate environmental input for decision-making. Combining the above multi-source information, the agent uses a large model for deep logical reasoning, ultimately outputting a quantified attitude score (e.g., [-10, 10]) and specific social response behaviors (e.g., posting, forwarding, replying, or remaining silent). This process transforms internal implicit cognitive states into externally observable explicit social signals, driving subsequent group interactions. Simultaneously, the simulator aggregates the attitude scores of all agents at the current time step in real time, calculating the average attitude index of the entire society through weighted averaging.

[0052] Step 4: To break the limitations of static networks and simulate group interaction and cognitive evolution based on news topic resonance in real-world event scenarios, this invention implements a dynamic clustering mechanism. Agents interested in the same news topic are temporarily aggregated into information-driven groups, constructing a local event field based on a specific news topic. During intra-group interactions, the simulator introduces a dynamic link prediction algorithm based on recommendation scores. This algorithm does not use random matching but comprehensively considers content alignment (measuring the semantic similarity of opinions and behaviors between agents) and timeliness factors (following the time decay law) to calculate recommendation scores between each agent. This mechanism guides agents to prioritize establishing connections with like-minded and highly active objects, realistically simulating the connection characteristics of small worlds in social networks and the natural decay process of trending news topics. This provides a path-dependent cognitive foundation for decisions in the next time step, thus achieving a logical closed loop from micro-interaction to macro-attitude evolution.

[0053] Step 5: After completing a round of decision-making and interaction (posting, replying, or forwarding), the agent triggers a self-reflection mechanism. Using a large model, it performs a deep summary and reflection on the current interaction experience (including its own behavior and observed neighbor feedback), internalizing instantaneous social interactions into long-term structured experience. This updates the memory and belief system, ensuring cognitive coherence in long-term simulations. Subsequently, the next phase of dual-stream data (containing new macro-events and news topic libraries) is loaded, while the agent's updated memory state is retained. Through this continuous iterative cycle of "perception-action-reflection," the simulator ultimately generates a trajectory strictly aligned with real-world history, completing the long-term social situation projection.

[0054] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0055] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0056] Based on the description of the above method embodiments, the present invention also provides a system. The system may be a system that uses software (applications), modules, components, servers, clients, etc., using the methods described in the embodiments of this specification, combined with necessary implementation hardware. Since the implementation schemes and methods for solving the problem are similar, the specific system implementations in the embodiments of this specification can be found in the implementations of the foregoing methods, and repeated details will not be elaborated upon.

[0057] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0058] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0059] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-agent cognitive self-evolution method, characterized in that, Includes the following steps: S1 initializes the global environment and agent group based on real-world data, and configures a person profile and dynamic memory for each agent; S2 constructs environmental information through dual-stream data consisting of macro event streams and news topic streams. It broadcasts the macro event streams as a global background to all agents and allows each agent to select the most relevant news topics from the news topic stream based on its own profile and dynamic memory. S3: The agent performs cognitive processing based on the received dual-stream data, user profile, and dynamic memory, and outputs attitude scores and behavioral responses. S4, based on behavioral response, dynamically clusters agents that are interested in the same news topic into temporary information-driven groups. Within the information-driven groups, agents are guided to interact dynamically based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memory and update dynamic memory. S5 aggregates the attitude scores and behavioral response data of all agents to synthesize the macro situation, uses the dynamic memory updated in the current round as the agent dynamic memory for the next round, and loads the dual-stream data for the next round. S6. Repeat steps S2 to S5 until the preset simulation termination condition is met.

2. The multi-agent cognitive self-evolution method according to claim 1, characterized in that, The feature allows each agent to select the most relevant news topics from the news topic stream based on its own profile and dynamic memory, specifically including: Constructing a candidate information pool based on news topic streams This enables the agent to react to macroscopic events in round t. Self-portrait And the dynamic memory of round t-1 It independently calculates and selects the most relevant news topics. : ； in, To select the engine function, the semantic alignment between the input news topics and the agent's persona, as well as dynamic memory, is comprehensively evaluated to retrieve the K news topics with the highest semantic alignment. This represents a large model.

3. The multi-agent cognitive self-evolution method according to claim 1, characterized in that, The intelligent agent performs cognitive processing based on the received dual-stream data, user profile, and dynamic memory, and outputs attitude scores and behavioral responses, specifically including: Introducing a decision engine based on prompt words With self-reflective memory operators In the t-th round of interaction, the i-th agent Attitude scores are generated based on input information. Behavioral response and the updated dynamic memory It will be made by a large model The decision-making process of a driven intelligent agent is modeled as follows: ； For the i-th agent, select the news topic that interests them most in round t. Create a portrait of the i-th intelligent agent. For the i-th agent, the dynamic memory in round t-1; The reference information flow is constructed by the i-th agent in the t-th round based on the visual reply content and dynamic information flow of the k-th news topic; The input information is composed of the above; The attitude score is used to quantify the agent's emotional polarity and emotional intensity towards current macro events and specific news topics; The behavioral response refers to the social response actions taken by the agent, including posting opinions, forwarding, or replying.

4. The multi-agent cognitive self-evolution method according to claim 1, characterized in that, The process of guiding agents to dynamically interact within an information-driven group based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memories and update dynamic memories, specifically includes: Based on visual response content With dynamic information flow Constructing a reference information flow ; Information-driven group for the k-th news topic in the t-th round of interaction In the process, calculate the i-th agent in the t-th round of interaction. With the i-th intelligent agent Recommended scores between For intelligent agents Sort the other agents in descending order of their recommendation scores and select the one with the highest recommendation score. individual agents and Enable dynamic interaction: ； in, Represents cosine similarity. The attenuation rate; Indicates the timestamp of the action's publication. This represents the action taken by the i-th agent in the t-th round of interaction regarding the current news topic. This represents the behavioral response of the j-th agent to the current news topic in the t-th round of interaction; Dynamic memory is updated based on reflective memory: ； This represents the dynamic memory of the i-th agent in round t. The prompt words are used to guide the large model in updating its memory. express After After processing and Dynamic memory updated jointly.

5. The multi-agent cognitive self-evolution method according to claim 4, characterized in that, The reference information stream The construction process specifically includes: Visual reply content This represents what the i-th agent can see in round t, targeting The set of observable responses: ； in, This represents the k-th news topic of interest selected by the i-th agent in the t-th round. This indicates that for the same news topic, the j-th agent... Execution with the i-th intelligent agent The same responsive action; Represents the i-th intelligent agent In the t round, targeting the A news item is a dynamic information stream received from other agents, which is obtained by aggregating all information streams from agents within the same group that perform the same operation driven by the same information: 。 6. A multi-agent cognitive self-evolutionary system, characterized in that, Includes the following steps: The initialization module initializes the global environment and agent group based on real-world data, and configures a person profile and dynamic memory for each agent. The dual-stream data integration module constructs environmental information through dual-stream data consisting of macro event streams and news topic streams. It broadcasts the macro event streams as a global background to all intelligent agents and allows each intelligent agent to select the most relevant news topics from the news topic stream based on its own profile and dynamic memory. In the initial behavior decision-making module, the agent performs cognitive processing based on the received dual-stream data, user profile, and dynamic memory, and outputs attitude scores and behavioral responses. The interactive reflection module, based on behavioral response, dynamically clusters agents that are interested in the same news topic into temporary information-driven groups. Within the information-driven groups, agents are guided to interact dynamically based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memories and update dynamic memories. The situation synthesis module aggregates the attitude scores and behavioral response data of all agents to synthesize the macro situation, uses the dynamic memory updated in the current round as the agent dynamic memory for the next round, and loads the dual-stream data for the next round.

7. A multi-agent cognitive self-evolutionary system according to claim 6, characterized in that, The feature allows each agent to select the most relevant news topics from the news topic stream based on its own profile and dynamic memory, specifically including: Constructing a candidate information pool based on news topic streams This enables the agent to react to macroscopic events in round t. Self-portrait And the dynamic memory of round t-1 It independently calculates and selects the most relevant news topics. : ； in, To select the engine function, the semantic alignment between the input news topics and the agent's persona, as well as dynamic memory, is comprehensively evaluated to retrieve the K news topics with the highest semantic alignment. This represents a large model.

8. A multi-agent cognitive self-evolutionary system according to claim 6, characterized in that, The intelligent agent performs cognitive processing based on the received dual-stream data, user profile, and dynamic memory, and outputs attitude scores and behavioral responses, specifically including: Introducing a decision engine based on prompt words With self-reflective memory operators In the t-th round of interaction, the i-th agent Attitude scores are generated based on input information. Behavioral response and the updated dynamic memory It will be made by a large model The decision-making process of a driven intelligent agent is modeled as follows: ； For the i-th agent, select the news topic that interests them most in round t. Create a portrait of the i-th intelligent agent. For the i-th agent, the dynamic memory in round t-1; The reference information flow is constructed by the i-th agent in the t-th round based on the visual reply content and dynamic information flow of the k-th news topic; The input information is composed of the above; The attitude score is used to quantify the agent's emotional polarity and emotional intensity towards current macro events and specific news topics; The behavioral response refers to the social response actions taken by the agent, including posting opinions, forwarding, or replying.

9. A multi-agent cognitive self-evolutionary system according to claim 6, characterized in that, The process of guiding agents to dynamically interact within an information-driven group based on content similarity and timeliness factors, triggering a self-reflection mechanism to generate reflective memories and update dynamic memories, specifically includes: Based on visual response content With dynamic information flow Constructing a reference information flow ; Information-driven group for the k-th news topic in the t-th round of interaction In the process, calculate the i-th agent in the t-th round of interaction. With the i-th intelligent agent Recommended scores between For intelligent agents Sort the other agents in descending order of their recommendation scores and select the one with the highest recommendation score. individual agents and Enable dynamic interaction: ； in, Represents cosine similarity. The attenuation rate; Indicates the timestamp of the action's publication. This represents the action taken by the i-th agent in the t-th round of interaction regarding the current news topic. This represents the behavioral response of the j-th agent to the current news topic in the t-th round of interaction; Dynamic memory is updated based on reflective memory: ； This represents the dynamic memory of the i-th agent in round t. The prompt words are used to guide the large model in updating its memory. express After After processing and Dynamic memory updated jointly.

10. A multi-agent cognitive self-evolutionary system according to claim 9, characterized in that, The reference information stream The construction process specifically includes: Visual reply content This represents what the i-th agent can see in round t, targeting The set of observable responses: ； in, This represents the k-th news topic of interest selected by the i-th agent in the t-th round. This indicates that for the same news topic, the j-th agent... Execution with the i-th intelligent agent The same responsive action; Represents the i-th intelligent agent In round t, targeting the A news item is a dynamic information stream received from other agents, which is obtained by aggregating all information streams from agents within the same group that perform the same operation driven by the same information: 。

Images